Unbreak random_agent

[avital]

The semantics of env.reset() have changed. It's no longer possible
to reset an environment until it's done.

Notably, this commit changes the behavior of random_agent. It used
to have a limit on number of actions per episode. If we wanted
that now, we'd have to close the environment and recreate it
on each episode, which may be slow.

[tlbtlbtlb]

LGTM

[avital]

Tests are failing, almost certainly not due to this change. Are there flaky tests?

FAIL: gym.envs.tests.test_envs_semantics.test_env_semantics:138
EnvSpec(Blackjack-v0)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/usr/local/gym/gym/envs/tests/test_envs_semantics.py", line 82, in test_env_semantics
    assert rollout_dict[spec.id]['rewards'] == rewards_now, 'Rewards not equal for {}'.format(spec.id)
AssertionError: Rewards not equal for Blackjack-v0